Is anyone else absolutely fed up—and I mean fed up—with ChatGPT/OpenAI API’s relentless use of em dashes—those long, obnoxious, totally un-human-like dashes—in its output? I just want human-sounding text—not something that screams “AI wrote this!”—and these dashes—they’re the first clue.
Seriously—who even uses these things all the time—besides an AI, that is? I don’t know how to type them—you probably don’t either—so why does the AI insist on them? They make everything feel so artificial—so cold—so robotic.
I’ve tried everything—prompting it to stop—asking it to use hyphens instead—begging it, practically—but the em dashes just keep coming—like a never-ending punctuation nightmare.
Has anyone—anyone at all—figured out a trick—a prompt—a secret code—to banish these hideous em dashes forever—and make the AI write like a real person—for once? My sanity—and my desire for natural AI output—depends on it! Help—please—I’m desperate!
The question I posted is an example of excessive em dashes. No matter what I put into the system prompt or even the model reminder, it still puts them in.
Not as bad as my question, that was done for emphasis. But I want NO DASHES.
My prompt says:
Guardrails & Boundaries
Never use em dashes
and the Model Reminder says :do not use em dashes. If you’re writing a post or message and want a pause, use a period or comma instead.
The closest I’ve come to getting an LLM to write like me is by using Claude and training it on my existing writing. I find it easier to get ChatGPT to ideate something, then tweak and edit it before sending it through 3.7 Sonnet for another draft.
The em dashes are an issue, but ChatGPT has lots of other tell-tale signs in its output like: “It’s not x, not y, just z.”
If you can, try using another model and seeing what the results are like.
It’s interesting you picked up on the not this but this framing; I have noticed that as well and it does seem to be pretty good at fixing it on the edit but no amount of prompting seems to truly eliminate it.
That’s exactly what I’m doing. I’m building a content creator, and it’s Claude all the way. It does a pretty good rendition of my voice, then I take the content out and polish it.
I’ve not tested this, but I ran it through a prompt writer I have and it’s given this:
Please avoid using em dashes (—) in your responses. Instead, wherever you would normally use an em dash, intelligently replace it with a comma, period, or semicolon based on the grammatical and stylistic needs of the sentence. Your priority is to preserve the intended emphasis or pause while using only those three forms of punctuation. Review your output to ensure that no em dashes are used.
The problem is, the use of em dashes is considered proper in nearly all forms of professional writing. So your prompt is going against nearly every book and published article the LLM has consumed. But I definitely feel your frustration - I use dashes like this and would prefer the output to do similarly. Especially since one program I often use AI content for (Tailwind, for Pinterest pins) will turn em dashes into its markdown code, which of course looks terrible.
You are correct. And all the formal writing in the training data, especially the older works, use em dashes a lot. But in modern web writing, almost no one uses em dashes. If someone has a – in a tweet or email you know it was AI drafted for sure.
Guardrails & Boundaries